Intelligent operation and maintenance monitoring system and method for arm computing power node

Abstract

The application belongs to the technical field of data monitoring, and discloses an intelligent operation and maintenance monitoring system and method for an ARM computing power node; comprising: collecting operation state data of the ARM computing power node and performing data cleaning to obtain node operation state data; performing node time synchronization processing on the node operation state data to generate node synchronization operation data; optimizing an alarm mechanism based on the node synchronization operation data to obtain adaptive alarm rules; executing abnormal response identification on the node synchronization operation data to output an abnormal response data queue; performing communication state detection on the abnormal response data queue to obtain communication state marking data; generating a node operation response strategy and sending the node operation response strategy to a preset ARM computing power node control terminal; the adaptive capacity and processing accuracy of abnormal information in the computing power network are significantly improved, and stronger operation and maintenance adaptability and service continuity guarantee are provided.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data monitoring, more particularly, the present application relates to an intelligent operation and maintenance monitoring system and method for ARM computing power nodes. BACKGROUND

[0002] With the development of new generation information technologies such as cloud computing and computing power network, the running architecture composed of ARM computing power nodes is widely deployed in various business scenarios such as cloud gaming, cloud mobile phone and mobile office due to its high energy efficiency and strong scalability. In order to ensure service continuity and stability, an efficient and intelligent operation and maintenance monitoring system is increasingly becoming a necessary component, enabling the running architecture composed of ARM computing power nodes to have the ability to perceive and respond to various abnormal conditions of computing power nodes. However, there are still some problems in existing monitoring technology.

[0003] Since the ARM computing power node runs instances on a computing board as a physical carrier, there is a problem of large difference in hardware structure, and the time reference of different boards depends on their own local crystal oscillator or clock module. Even if a unified sampling trigger mechanism is used when collecting node running data, it is difficult to eliminate the misalignment of sampling time stamps between multiple nodes, and it is not suitable for application to cloud platforms which require second-level response of the operating system. In terms of alarm mechanism, the existing technology often lacks dynamic perception of the current running task type, leading to misjudgment or missed judgment, for example, high CPU or GPU utilization may be normal in some task types, but may indicate a computing bottleneck or delay risk when running another task type. In addition, if a large number of state indicators are reported by the node during running, the receiving terminal has difficulty in identifying the core abnormal indicators that really need to be processed quickly, resulting in key signals being submerged or even ineffective, leaving potential safety hazards. At the same time, in terms of communication detection, the traditional technical solution often uses ping or port response to judge whether the communication state is normal, but the semantic content in the response packet or resource writing defects may easily lead to false communication failure, at this time the content is abnormal but the communication link structure may not be abnormal, and the system will directly judge it as communication failure due to identification timeout or response format abnormality, resulting in misjudgment of the communication state. When the system determines that an abnormality has occurred, it will execute the corresponding abnormality handling strategy or restart the system to correct the abnormality. However, if the state indicators repeatedly fluctuate around the system's default abnormal threshold, the system may repeatedly restart or execute the abnormality handling strategy, resulting in a substantial increase in system resource overhead, which seriously threatens the stability of the service.

[0004] In view of this, the present application proposes an intelligent operation and maintenance monitoring system and method for ARM computing power nodes to solve the above problems. SUMMARY

[0005] In order to overcome the above-mentioned defects of the prior art, in order to achieve the above-mentioned purpose, the present application provides the following technical scheme: an intelligent operation and maintenance monitoring method of an ARM computing power node, comprising:

[0006] S1. Collecting the running state data of the ARM computing power node and performing data cleaning to obtain node running state data;

[0007] S2. Performing node time synchronization processing on the node running state data to generate node synchronization running data;

[0008] S3. Optimizing the alarm mechanism based on the node synchronization running data to obtain adaptive alarm rules;

[0009] S4. Combining the adaptive alarm rules to execute abnormal response identification on the node synchronization running data, and outputting an abnormal response data queue;

[0010] S5. Detecting the communication state of the abnormal response data queue to obtain communication state marking data;

[0011] S6. Generating a node running response strategy based on the communication state marking data, and sending the node running response strategy to a preset ARM computing power node control terminal.

[0012] Further, the node time synchronization processing method comprises:

[0013] Obtaining the board card number to which the ARM computing power node belongs and grouping the node running state data to output grouped node data; setting a global sampling time standard of the node unit; calculating the difference between the timestamp of the grouped node data and the global sampling time standard of the node unit to obtain the time offset of the corresponding board card node unit; comparing the time offset with a preset time error tolerance threshold, and dividing the node running state data into a synchronous node data subset and a non-synchronous node data subset according to the comparison result;

[0014] Extracting features from the non-synchronous node data subset to obtain node time features; performing time series clustering processing on the non-synchronous node data subset based on the node time features to generate a set of lag node clusters and a set of drift trend node clusters; obtaining historical sampling response data of different board cards, and performing time compensation mapping on the set of lag node clusters based on the historical sampling response data to output a set of time corrected node clusters; extracting the drift period of the set of drift trend node clusters and calculating the drift rate, and performing time axis reparameterization processing on the set of drift trend node clusters based on the drift period and the drift rate to obtain a set of time reconstructed node clusters; synchronizing the synchronous node data subset, the set of time corrected node clusters, and the set of time reconstructed node clusters to obtain the node synchronization running data.

[0015] Further, the optimization alarm mechanism comprises:

[0016] obtain a running task identifier corresponding to the node synchronous running data in real time, match the running task identifier with the node number, and obtain a task node index set; determine a task type based on the task node index set, and construct a resource allocation indicator of each task type; extract historical node data of the corresponding task type, calculate a multi-dimensional parameter difference between the resource allocation indicator of the historical node data and the resource allocation indicator of the current node, perform sensitivity weight allocation on the resource allocation indicator based on the multi-dimensional parameter difference, and output a weighted resource allocation indicator;

[0017] perform interactive correlation analysis on the weighted resource allocation indicator, and output an indicator fluctuation correlation feature; construct a task type abnormality judgment condition in combination with the indicator fluctuation correlation feature and the task type; establish a mapping relationship between the task type abnormality judgment condition and the weighted resource allocation indicator, construct an abnormal correlation data set based on the mapping relationship; identify an abnormal subset of task overlap in the abnormal correlation data set, generate an alarm intensity sequence based on a trigger response sequence of related indicators in the abnormal subset; and combine the related indicators, the task type abnormality judgment condition, and the alarm intensity sequence to form an adaptive alarm rule.

[0018] Further, the manner of constructing the task type abnormality judgment condition comprises:

[0019] The task type includes cloud gaming, mobile office, and cloud phone;

[0020] When the task type is cloud gaming, the task type abnormality judgment condition includes GPU short-time utilization rate change, continuous frame rate jitter frequency, and image rendering delay;

[0021] When the task type is mobile office, the task type abnormality judgment condition includes CPU mutation change rate, memory allocation, and work task switching frequency;

[0022] When the task type is cloud phone, the task type abnormality judgment condition includes system access delay, network delay jump frequency, and data interruption frequency.

[0023] Further, the manner of performing abnormal response identification comprises:

[0024] Take the adaptive alarm rule as a judgment basis to identify abnormal response conditions in the node synchronous running data, and output preliminary screening abnormal data; identify running behaviors corresponding to the preliminary screening abnormal data, and screen abnormal processing subsets based on the running behaviors; set an abnormal danger level for each abnormal processing subset, and determine related abnormal indicators with an abnormal danger level higher than a preset danger level threshold as preferred abnormal indicators;

[0025] The dynamic sliding window traverses all the abnormal handling subsets to obtain all the preferred abnormal indicators in the dynamic sliding window at the same time, and determines the preferred abnormal indicators as the highest processing priority.

[0026] For non-preferred abnormal indicators, the distribution density and duration of the non-preferred abnormal indicators are obtained by using the dynamic sliding window at the same time; the load state features of the corresponding ARM computing power nodes are collected in real time, and the sub-optimal abnormal score is calculated by combining the distribution density, duration and load state feature values; the non-preferred abnormal indicators are sorted based on the sub-optimal abnormal score to obtain a sub-optimal abnormal queue; and the preferred abnormal indicators and the sub-optimal abnormal queue are integrated to obtain an abnormal response data queue.

[0027] Further, the communication state detection method comprises:

[0028] The communication path set of the ARM computing power node corresponding to the abnormal response data queue is obtained; the node response data of each communication path in the communication path set is extracted, and the semantic content of the node response data is identified, and a node semantic vector matrix is constructed based on the semantic content; the structural integrity of the node semantic vector matrix is analyzed, and the initial abnormal response data packet is screened based on the structural integrity analysis result;

[0029] The request instruction corresponding to the remaining node response data is identified, and the original containing field is queried based on the request instruction; the mismatch rate is calculated by comparing the number of existing fields and the original containing field, and the remaining node response data is classified into two categories based on the mismatch rate and the preset threshold to obtain a two-class abnormal response data packet;

[0030] The resource occupation rate of the remaining node response data is extracted, and an abnormal field occupation vector is constructed; the standard deviation of all elements in the abnormal field occupation vector is calculated, the skewness of the abnormal field occupation vector is analyzed based on the standard deviation, and the skewness coefficient is output; if the skewness coefficient does not conform to the preset standard coefficient interval, the remaining node response data is classified into three categories to obtain a three-class abnormal response data packet; and the communication abnormal state label of the abnormal response data packet obtained by each screening is marked to obtain a communication state label data.

[0031] Further, the communication abnormal state labeling method comprises:

[0032] If the structural integrity analysis result is structural incompleteness, the initial abnormal response data packet at this time is determined as a structural damage abnormal state; if the mismatch rate is higher than the preset threshold, the two-class abnormal response data packet at this time is determined as a semantic content abnormal state; and if the skewness coefficient does not conform to the preset standard coefficient interval, the three-class abnormal response data packet at this time is determined as a resource write abnormal state.

[0033] Further, the method for generating a node running response strategy comprises:

[0034] extracting a communication state label of the communication state marker data, and matching the communication state label and the related abnormality indicators of the corresponding data with a preset abnormality solution strategy database to generate a suitable strategy instruction;

[0035] extracting a fluctuation value of the related abnormality indicators of any one node in a continuous preset monitoring period, and constructing a state fluctuation matrix based on the value; setting a fluctuation reasonable interval, and counting the number of times that all related abnormality indicator fluctuation values of each node in the state fluctuation matrix exceed the fluctuation reasonable interval; if the number of times is greater than a preset frequency threshold, an abnormality processing action is directly executed; otherwise, the corresponding node is temporarily marked, and the trend change of the related abnormality indicators of the temporarily marked node in the subsequent continuous period is recorded to construct a trend deviation function; if the value of the trend deviation function does not exceed a preset accumulation threshold and the change direction is correct in the continuous period, the temporary mark is eliminated and the abnormality processing action is abandoned; if the value of the trend deviation function continuously exceeds a preset function threshold and the change direction is an upward trend in the continuous period, the abnormality processing action is directly executed; and the suitable strategy instruction and the encoded abnormality processing action execution process are integrated to construct a node operation response strategy.

[0036] Further, the state fluctuation matrix is constructed in the following manner:

[0037] all related abnormality indicators belonging to the same node are normalized to obtain normalized abnormality indicators; the normalized abnormality indicators are spliced into a row vector, and the row vectors of the normalized abnormality indicators corresponding to all nodes are superimposed to obtain the state fluctuation matrix.

[0038] An intelligent operation and maintenance monitoring system of an ARM computing power node is used to implement an intelligent operation and maintenance monitoring method of the ARM computing power node, and the system is characterized in that the system comprises:

[0039] a data acquisition module configured to acquire operation state data of the ARM computing power node and perform data cleaning to obtain node operation state data;

[0040] a node synchronization module configured to perform node time synchronization processing on the node operation state data to generate node synchronization operation data;

[0041] an alarm optimization module configured to optimize an alarm mechanism based on the node synchronization operation data to obtain an adaptive alarm rule;

[0042] an abnormality identification module configured to execute abnormality response identification on the node synchronization operation data in combination with the adaptive alarm rule to output an abnormality response data queue;

[0043] a communication detection module configured to perform communication state detection on the abnormality response data queue to obtain communication state marker data;

[0044] The strategy generation module is used to generate node operation response strategies based on communication status marker data and send the node operation response strategies to the preset ARM computing power node control terminal; the modules are connected to each other via wired and / or wireless means.

[0045] The technical effects and advantages of the intelligent operation and maintenance monitoring system and method for ARM computing nodes of this invention are as follows:

[0046] By collecting operational status data from ARM nodes, this paper optimizes aspects such as node synchronization, alarm mechanisms, anomaly response, communication detection, and policy generation, thus realizing an intelligent operation and maintenance monitoring method for ARM computing nodes. Compared with existing experience, the introduction of a multi-node time synchronization process effectively solves the problem of clock inconsistency caused by different board hardware structures, enabling the system to adapt to second-level response in application scenarios. By integrating task type and resource sensitivity, the system can adjust alarm mechanism thresholds according to the actual business impact of tasks, avoiding false alarms or missed alarms that are prone to occur in traditional alarm rules. Furthermore, the introduction of semantic content detection and skewness analysis... This method breaks through the traditional communication detection technology's judgment method of "a response means successful communication, otherwise communication fails," effectively mitigating the situation of misjudgment of anomalies and improving the accuracy of communication anomaly location. By detecting the fluctuation trend of node indicators, it realizes anti-jitter control during the operation of computing power nodes, effectively avoiding the high-overhead operation of the system to execute anomaly response strategies due to repeated slight fluctuations in node-related indicators, thus improving the overall operational stability. In summary, this intelligent operation and maintenance monitoring method for ARM computing power nodes significantly improves the adaptability and processing accuracy of node anomaly information in the computing power network, providing stronger operation and maintenance adaptability and service continuity assurance. Attached Figure Description

[0047] Figure 1 This is a schematic diagram of an intelligent operation and maintenance monitoring method for ARM computing nodes according to the present invention;

[0048] Figure 2 This is a schematic diagram of an intelligent operation and maintenance monitoring system for ARM computing nodes according to the present invention. Detailed Implementation

[0049] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0050] Example 1

[0051] Please see Figure 1As shown, the intelligent operation and maintenance monitoring method of the ARM computing power node in the embodiment comprises:

[0052] S1. Collecting the running state data of the ARM computing power node and performing data cleaning to obtain node running state data;

[0053] S2. Performing node time synchronization processing on the node running state data to generate node synchronization running data;

[0054] S3. Optimizing the alarm mechanism based on the node synchronization running data to obtain adaptive alarm rules;

[0055] S4. Combining the adaptive alarm rules to execute abnormal response identification on the node synchronization running data, and outputting an abnormal response data queue;

[0056] S5. Detecting the communication state of the abnormal response data queue to obtain communication state marking data;

[0057] S6. Generating a node running response strategy based on the communication state marking data, and sending the node running response strategy to a preset ARM computing power node control terminal.

[0058] The application scenario of the embodiment is 100 sets of M114 Android cloud server clusters, and each set of M114 contains 84 ARM computing board cards, each board card corresponds to an independent virtualized Android instance, which is used to provide ARM computing power services, and the ARM computing board card is the physical carrier output by the ARM computing power node; wherein the running state data includes physical parameters reflecting the node running condition, such as CPU usage, memory occupancy, process survival state and GPU rendering queue length; at the same time, the running state data is first filled with missing values and compressed with redundant data in the embodiment, data cleaning is realized, and higher quality node running state data is obtained.

[0059] Since each server cluster has 84 ARM computing board cards, a high-density heterogeneous physical structure is formed; each board card is essentially a small computing subsystem, and its time reference depends on its local crystal oscillator or BMC clock module; and in different specific application scenarios such as cloud gaming or mobile office, different nodes will receive corresponding business behavior restrictions, so even if uniform sampling time is used when collecting, small time deviations cannot be avoided, which has a significant impact on businesses that require second-level responses.

[0060] The way of performing node time synchronization processing includes:

[0061] The board card number to which the ARM computing power node belongs is acquired, and the node running state data is grouped, and the grouped node data is output. The running instance in the ARM computing power node is mapped to the ARM board card number where it is deployed, and grouping is performed based on the number. Each group of grouped node data belongs to the same ARM computing board card, ensuring clear division of data.

[0062] A node unit global sampling time standard is set. The node unit global sampling time standard is a unified time reference line preset by the control terminal of the ARM computing power node in actual application, and is used for subsequent time synchronization processing operation. The difference between the timestamp of the grouped node data and the node unit global sampling time standard is calculated to obtain the time offset of the corresponding board card node unit. The time offset is used to measure the time difference between the sampling timestamp of the grouped node data corresponding to each board card and the set node unit global sampling time standard. The difference between the sampling timestamp of different board cards and the unified time standard is used to reflect the degree of time offset.

[0063] The time offset is compared with a preset time error tolerance threshold, and the node running state data is divided into a synchronous node data subset and a non-synchronous node data subset according to the comparison result. The time error tolerance threshold is a threshold set based on the historical experience of those skilled in the art, which is used to determine whether the time offset meets the threshold requirement and classify the nodes. When the time offset exceeds the preset time error tolerance threshold, the corresponding node unit is determined as a non-synchronous node and combined into a non-synchronous node data subset. Otherwise, the node unit is determined as a synchronous node and combined into a synchronous node data subset.

[0064] Feature extraction is performed on the non-synchronous node data subset to obtain node time features. The node time features include, for example, sampling interval fluctuation, sampling order stability, and offset trend of the corresponding node, and are used to reflect the change behavior trajectory of the non-synchronous node data subset.

[0065] Based on the node time features, the non-synchronous node data subset is subjected to time series clustering processing to generate a set of lag node clusters and a set of drift trend node clusters. In this embodiment, the DTW dynamic time warping algorithm is used to identify data with similar time lag behavior in the non-synchronous node data subset and cluster them into a set of lag node clusters. Data with similar increasing or decreasing trend of continuous offset are clustered into a set of drift trend node clusters, which are used to reflect two abnormal sampling modes in the non-synchronous node data subset.

[0066] The historical sampling response data of different boards is acquired, and time compensation mapping is performed on the hysteresis node cluster set based on the historical sampling response data, and a time correction node cluster set is output, wherein the historical sampling response data of a board is obtained by extracting the sampling timestamp data of the board in a plurality of previous sampling periods; the nodes with response lag in the historical sampling response data are identified, and it is determined whether these lag nodes exist in continuous sampling periods, if yes, it is indicated that these nodes are stable hysteresis, and the time compensation mapping process is entered; the compensation value is obtained by calculating the average lag time of the response lag nodes in the historical sampling response data, and the value is used for time compensation mapping of the current hysteresis node corresponding to the board; if it is not stable hysteresis, it is determined as temporary hysteresis, and the sampling timestamp of the node is aligned with the global sampling time standard of the node unit to realize compensation; and the time correction node cluster set is obtained by integrating all the compensated hysteresis nodes.

[0067] The drift period of the drift trend node cluster set is extracted, and the drift rate is calculated, time axis reparameterization processing is performed on the drift trend node cluster set based on the drift period and the drift rate, and a time reconstruction node cluster set is obtained, wherein the drift trend node cluster set is different from the hysteresis node cluster set, the offset of the hysteresis node cluster set is small and can be compensated, and the time offset value of the drift trend node cluster set is in rapid change in a certain trend, so the time axis reparameterization method is used for processing in this embodiment.

[0068] The time axis reparameterization is a technology widely used in many fields for processing time series data; the drift period of the drift trend node cluster set is obtained, and the drift rate is calculated, and then the drift period and the drift rate are used as the basic variables of the reparameterization function in the time axis reparameterization method; it should be noted that the drift period is obtained as follows: a waveform graph of a node cluster is drawn based on the node values of each node cluster in the drift trend node cluster set, wherein the horizontal axis represents time, and the vertical axis represents the specific value of the node in the node cluster; the time interval of adjacent repeated offset waveforms is extracted from the waveform graph, which is the drift period; the drift rate refers to the slope of the sampling point in the offset waveform in any drift period; the drift period is used to guide the period fitting and stability setting of the reparameterization function, and the drift rate is used to guide the slope direction of the reparameterization function for adjusting the sampling time; finally, the offset amount of the drift trend node cluster that needs to be compensated is calculated by using the reparameterization function, and the time reconstruction node cluster set is obtained; the node synchronous operation data is obtained by integrating the synchronous node data subset, the time correction node cluster set and the time reconstruction node cluster set.

[0069] The manner for optimizing the alarm mechanism includes:

[0070] The running task identifier corresponding to the real-time node synchronization running data is obtained, and the running task identifier is matched with the node number to obtain a task node index set. The running task identifier refers to the task being performed by the relevant ARM computing power node of the node synchronization running data at the time of use, such as a specific running task of running a game, browsing a webpage, or establishing communication. Each ARM computing power node carried by each board has a number, so each specific running task is bound to the corresponding node number, and the task node index set is obtained.

[0071] The task type is determined based on the task node index set, and a resource allocation indicator of each task type is constructed. Based on the specific running task in the task node index set, the general type of the task is divided into three categories: cloud gaming, mobile office, and cloud phone. For each task type, the corresponding resource parameters consumed in the historical stable running state are extracted, and the corresponding resource parameters consumed by each task type are taken as the resource allocation indicator of the node belonging to the task type.

[0072] The historical node data of the corresponding task type is extracted, the multi-dimensional parameter difference between the resource allocation indicator of the historical node data and the resource allocation indicator of the current node is calculated, the sensitivity weight of the resource allocation indicator is allocated based on the multi-dimensional parameter difference, and the weighted resource allocation indicator is output. The multi-dimensional parameter difference refers to the difference between each dimension parameter of the resource allocation indicator of the node data in the historical stable state and the resource allocation indicator of the current node, which is used to measure the deviation degree between the current running state and the historical stable state. The calculated multi-dimensional parameter difference is normalized to eliminate the dimensional difference, and the dimensionless characteristic value belonging to the resource allocation indicator is obtained. The initial sensitivity weight corresponding to each task type is set, and the initial sensitivity weight is set by querying the historical record, and the initial sensitivity weight is adjusted based on the specific task situation. The weighted resource allocation indicator of a certain task type is obtained by multiplying the sensitivity weight and the dimensionless characteristic value, which avoids the false alarm situation caused by the fixed threshold.

[0073] The weighted resource allocation indicator is interactively analyzed, and the indicator fluctuation correlation feature is output. The interactive correlation analysis is used to identify the coupling behavior relationship that may exist between individual indicator parameters in each weighted resource allocation indicator. The indicator fluctuation correlation feature refers to the correlation strength of multiple coupling behaviors, such as the frame rate jitter amplitude accompanied by the increase of GPU occupancy, the system transmission efficiency change caused by memory criticality, etc. The task type abnormality judgment condition is constructed by combining the indicator fluctuation correlation feature and the task type. By matching the indicator fluctuation correlation feature with the task type, the abnormality judgment condition of the corresponding task type is constructed, that is, a certain combination of indicators can constitute an effective alarm triggering condition under a certain task type.

[0074] The mapping relationship between the task type abnormality judgment condition and the weighted resource allocation index is established, and the abnormal association data set is constructed based on the mapping relationship, wherein the abnormality judgment condition and the weighted resource allocation index corresponding to the task type of the condition are bound, and the corresponding mapping relationship is established; the abnormal association data set of multiple abnormal events with mutual association relationship is constructed in combination with the mapping relationship.

[0075] The abnormal subset of task overlap in the abnormal association data set is identified, and the alarm intensity sequence is generated based on the trigger response sequence of the related indexes in the abnormal subset, wherein several abnormal events with association relationship are extracted as the abnormal subset, and the trigger response sequence refers to which related index enters the abnormality judgment state first, and the related indexes are sorted in descending order based on the sequence; it should be noted that if several related indexes enter the abnormality judgment state at the same time, the related indexes are sorted in descending order based on the sensitivity weight size at this time, and the alarm intensity sequence is finally obtained by integrating all sorted related indexes; the adaptive alarm rule is constructed in combination with the related indexes, the task type abnormality judgment condition and the alarm intensity sequence, the related indexes are corresponded to the task type abnormality judgment condition, and the sequence rule of the related indexes of multiple task types when performing abnormality judgment is constructed in combination with the alarm intensity sequence.

[0076] The construction method of the task type abnormality judgment condition includes:

[0077] The task types include cloud gaming, mobile office and cloud phone, and the use characteristics of system resources in the running process of different types of tasks are obviously different, and the sensitivity of different task types to resource changes is different.

[0078] When the task type is cloud gaming, the task type abnormality judgment condition includes GPU short-time utilization rate change, continuous frame rate jitter frequency and image rendering delay, since the essential logic of cloud gaming is a strong dependence application of real-time graphics rendering and interactive feedback, the change of GPU related indexes is sensitive, therefore, the GPU short-time utilization rate change, the continuous frame rate jitter frequency and the image rendering delay are taken as the abnormality judgment conditions and the corresponding thresholds are set; the GPU short-time utilization rate change can reflect the instantaneous rendering congestion caused by the surge of rendering load, the continuous frame rate jitter frequency can reflect whether the frame rate is stable in a period of continuous time; the image rendering delay can reflect whether the output display state is delayed to affect the game experience.

[0079] When the task type is mobile office, the task type exception determination conditions include CPU mutation change rate, memory allocation situation and work task switching frequency. When in the mobile office task type, the system is mainly for processing documents and light computing, generally not involving GPU level, so the determination conditions are focused on CPU and memory level and the corresponding threshold is set to determine whether the office task can respond smoothly; the CPU mutation change rate is used to detect whether there is unexpected CPU high-speed idling or burst scheduling failure; the memory allocation situation is used to reflect whether the distribution of the system in the office is reasonable; the work task switching frequency is used to detect whether there is task jam or multi-task congestion.

[0080] When the task type is cloud phone, the task type exception determination conditions include system access delay, network delay jump frequency and data interruption times. The running environment of cloud phone task type is more complex, which may be accompanied by a large number of behaviors such as simulated communication, background task and virtual event. The most sensitive core indicators of this task type are the response accessibility and communication stability of the system. Therefore, the system access delay, network delay jump frequency and data interruption times are used as determination conditions and the corresponding threshold is set; the system access delay is used to judge whether the system is running normally; the network delay jump frequency is used to judge the network stability; the data interruption times are used to measure whether there is abnormal interruption or response timeout phenomenon in the communication link.

[0081] The execution exception response recognition mode includes:

[0082] The adaptive alarm rule is used as the basis for identifying the abnormal response in the node synchronization running data, and the preliminary screening abnormal data is output. The data in the node synchronization running data is matched with the adaptive alarm rule. If it meets the conditions of the adaptive alarm rule, it is determined as abnormal data. All abnormal data is integrated to obtain preliminary screening abnormal data. The adaptive alarm rule is an early warning for possible abnormal data. The abnormal response recognition at this time is the real identification of the existing abnormal data.

[0083] The running behavior corresponding to the preliminary screening abnormal data is identified, and the abnormal processing subset is screened based on the running behavior. The behavior path analysis is performed on the running task to which the preliminary screening abnormal data belongs. First, it is judged which task type it belongs to. Then, the subset that needs to be processed abnormally is screened from the preliminary screening abnormal data according to the running behavior; the running behavior refers to the parameters that can reflect the system resource consumption at this time under the current task type, including process scheduling state and thread load.

[0084] An abnormal danger level is set for each abnormal handling subset, and an abnormal index related to the abnormal danger level higher than a preset danger level threshold is determined as an optimal abnormal index, wherein the preset danger level threshold refers to a threshold for distinguishing between general behavior and high-risk behavior based on expert experience and known theoretical knowledge, and if the abnormal danger level of a certain abnormal handling subset is higher than the threshold, it means that the running behavior corresponding to the abnormal handling subset belongs to high-risk behavior; the indexes in each abnormal handling subset are taken as units to determine the danger level of each index by querying historical records, and if there is an abnormal index related to the abnormal danger level higher than the preset danger level threshold, it is determined as an optimal abnormal index.

[0085] A dynamic sliding window is constructed to traverse all abnormal handling subsets, all optimal abnormal indexes in the dynamic sliding window at the same time are obtained, and the optimal abnormal indexes are determined as the highest processing priority, wherein the dynamic sliding window is used to divide the complete time range into several time periods by traversal, and the optimal abnormal indexes in different abnormal handling subsets in the same time period are identified; all the optimal abnormal indexes in the time period are determined as the highest processing priority.

[0086] For non-optimal abnormal indexes, the distribution density and duration of non-optimal abnormal indexes are obtained by using the dynamic sliding window at the same time, wherein the remaining indexes that are not optimal abnormal indexes are non-optimal abnormal indexes; the distribution density and duration of non-optimal abnormal indexes in the same time period as the optimal abnormal indexes are obtained, wherein the distribution density refers to the number of occurrences of a certain non-optimal abnormal index in the time period, and the duration indicates the cumulative duration of the intermittent or continuous occurrence of the index.

[0087] The load state features corresponding to the ARM computing power nodes are collected in real time, the sub-optimal abnormal score is calculated by combining the values of the distribution density, duration and load state features, the non-optimal abnormal indexes are sorted based on the sub-optimal abnormal score, and a sub-optimal abnormal queue is obtained, wherein the calculation formula of the sub-optimal abnormal score is: ; wherein, represents the sub-optimal abnormal score of any non-optimal abnormal index; represents the distribution density of the non-optimal abnormal index; represents the duration of the non-optimal abnormal index; represents the number of load state features; represents the mean value of the load state features; represents the standard deviation of the load state features; represents the value of the th load state feature, wherein the load state features include parameters such as CPU utilization, GPU rendering utilization and memory write blocking rate; , and respectively represent the weight of the sum of the distribution density, the duration and the normalized load characteristic value, which are set in this embodiment , , The skilled in the art can adjust according to the specific circumstances; the sub-optimal abnormal queue is obtained by descending arrangement based on the sub-optimal abnormal score.

[0088] The abnormal response data queue is obtained by integrating the preferred abnormal indicators and the sub-optimal abnormal queue, wherein the preferred abnormal indicators are marked as the highest processing priority, and the rest are arranged according to the sub-optimal abnormal queue, thereby obtaining the abnormal response data queue of a certain time period.

[0089] The communication state detection method comprises the following steps:

[0090] The communication path set of the ARM computing power node corresponding to the abnormal response data queue is obtained, wherein the communication path refers to the communication link connection established in the network connection in the specific running process of the ARM computing power node, and each communication path comprises information such as transmission protocol data chain, session channel and response port.

[0091] The node response data of each communication path in the communication path set is extracted, and the semantic content of the node response data is identified, and a node semantic vector matrix is constructed based on the semantic content, wherein the node response data refers to the response data returned along the communication path after the ARM computing power node receives the system instruction, which comprises information such as message structure, field content and status code; the semantic content of the node response data is extracted by using a preset semantic large model, and the semantic content is encoded into a vector form to construct a semantic vector matrix.

[0092] The structural integrity of the node semantic vector matrix is analyzed, and the initial abnormal response data packet is screened based on the structural integrity analysis result, wherein whether the fields of each data packet in the node semantic vector matrix are complete, whether the field order conforms to the communication protocol specification, and whether the data type can be identified are judged, if any one or more of the above requirements are not met, the corresponding node response data is determined as the initial abnormal response data packet.

[0093] The request instruction corresponding to the rest of the node response data is identified, and the original containing field is queried based on the request instruction, wherein the rest of the node response data refers to the data passing the structural integrity analysis except the initial abnormal response data packet; the original containing field is obtained by indexing and querying the field content returned by the corresponding request instruction in the initial request of the rest of the node response data.

[0094] The mismatch rate is calculated by comparing the existing field with the number of original included fields, and the remaining node response data is screened based on the mismatch rate and a preset threshold to obtain the second-class abnormal response data packet. The field matching rate is obtained by calculating the ratio of the actual returned field to the theoretical standard returned field number, and the mismatch rate is obtained by subtracting the field matching rate from 1. When the mismatch rate is higher than the preset threshold, it is determined as the second-class abnormal response data packet. The preset threshold is set by the person skilled in the art according to the actual demand.

[0095] The resource occupancy rate of the remaining node response data is extracted to construct an abnormal field occupancy vector. The remaining node response data is the same as the above-mentioned remaining node response data, and is also the data analyzed by the structural integrity except the initial abnormal response data packet. The resource occupancy rates of all nodes corresponding to the remaining node response data are combined into an array and converted into a vector form to obtain the abnormal field occupancy vector.

[0096] The standard deviation of all elements in the abnormal field occupancy vector is calculated, and the skewness analysis is performed on the abnormal field occupancy vector based on the standard deviation to output the skewness coefficient. The skewness analysis is a method for measuring the asymmetry of data distribution in statistics, which calculates the skewness coefficient by constructing a third-order standard moment formula by calling the standard deviation. The skewness coefficient is used to measure the degree of asymmetry of data distribution.

[0097] If the skewness coefficient does not meet the preset standard coefficient interval, the third-class screening is performed on the remaining node response data to obtain the third-class abnormal response data packet. The preset standard coefficient interval refers to the theoretical value interval of the skewness coefficient calculated by the skewness analysis in the current application scenario based on the theoretical knowledge of skewness analysis.

[0098] In this embodiment, if the skewness coefficient is greater than 0 and greater than or equal to the upper boundary of the preset standard coefficient interval, it means that the resource occupancy rate of most fields is normal, and a few fields may be 0 or NULL state, which may exist, for example, interface field null or resource field missing risk of resource not written; if the skewness coefficient is less than 0 and less than or equal to the lower boundary of the preset standard coefficient interval, it means that the resource occupancy rate of most fields is abnormal, and a few fields are normal, which may exist, for example, one-sided resource abnormal trend of most fields high occupancy. The communication abnormal state label of the abnormal response data packet obtained by each screening is obtained, and the communication state label data is obtained by marking the communication abnormal state corresponding to the abnormal response data packet obtained by each screening as a label. The communication state label data containing the communication abnormal state label is obtained.

[0099] The abnormal response data packets screened each time are: the initial abnormal response data packets screened according to the initial structural integrity analysis result; the second-type abnormal response data packets with semantic content abnormalities obtained after the second-type screening of the nodes other than the initial abnormal response data packets; and the third-type abnormal response data packets with abnormal resource writing obtained after the third-type screening of the nodes other than the initial abnormal response data packets. It should be noted that the nodes screened each time may appear in the second-type screening and the third-type screening. Therefore, the state superposition marking method is used for marking the communication abnormal state of the nodes, that is, one node may have two communication abnormal states.

[0100] The marking method of the communication abnormal state includes:

[0101] If the structural integrity analysis result is structural incompleteness, the initial abnormal response data packet is determined as a structural damage abnormal state. If the structural integrity analysis result is structural incompleteness, it indicates that there is a structural loss or other abnormal situation, and a structural damage abnormal state label is generated to mark the data. If the mismatch rate is higher than the preset threshold, the second-type abnormal response data packet is determined as a semantic content abnormal state. If the mismatch rate is higher than the preset threshold, it indicates that the field content is incorrect, and a semantic content abnormal state label is generated to mark the data. If the skewness coefficient does not meet the preset standard coefficient interval, the third-type abnormal response data packet is determined as a resource writing abnormal state. If the skewness coefficient cannot fall within the preset standard coefficient interval, it indicates that the resource occupation is unbalanced, and a resource writing abnormal state label is generated to mark the data.

[0102] The method for generating a node running response strategy includes:

[0103] The communication state label of the communication state marking data is extracted, and the communication state label and the related abnormal indicators corresponding to the data are matched with the preset abnormal solution strategy database to generate a suitable strategy instruction. The abnormal state corresponding to the communication state label and the abnormal state corresponding to the related abnormal indicators are used as preconditions, matched with the preset abnormal solution strategy database, and an instruction combination capable of solving the above abnormal state is generated, that is, a suitable strategy instruction.

[0104] The related abnormal indicator fluctuation value of any node in a continuous preset monitoring period is extracted, and a state fluctuation matrix is constructed based on the value. The related abnormal indicator fluctuation value refers to the change amplitude of a plurality of related abnormal indicators corresponding to a node in a continuous preset monitoring period. The abnormal state fluctuation matrix is used to horizontally reflect the fluctuation of each indicator of a node.

[0105] Set a reasonable fluctuation interval, count the number of times that the fluctuation value of each node in the state fluctuation matrix exceeds the reasonable fluctuation interval, wherein the reasonable fluctuation interval refers to the maximum deviation threshold that each index can accept under normal fluctuation state.

[0106] If the number of times is greater than the preset frequency threshold, directly execute the abnormal handling action, wherein the preset frequency threshold refers to a reasonable number of times limit value set by querying existing data and expert experience; if the change range of all related abnormal indexes of a certain node exceeds the reasonable fluctuation interval and the number of times is greater than the preset frequency threshold, the abnormal at this time is determined as high abnormal fluctuation, and an abnormal handling action capable of handling the abnormal situation at this time is selected from the preset abnormal solution strategy database to rapidly handle it.

[0107] Otherwise, mark the corresponding node as suspended, record the trend change of the related abnormal indexes of the suspended node in the subsequent continuous period, and construct a trend offset function, wherein if the number of times is less than the preset frequency threshold, the corresponding node may be in a slightly fluctuating state, and the system does not directly execute the abnormal handling action, but marks it as suspended, indicating that the node has potential abnormality but has not yet produced serious fluctuation out-of-control phenomenon, preventing false triggering of the abnormal handling action of the system and affecting normal operation; record the value change of the corresponding related abnormal indexes of the suspended node in the subsequent continuous period, and construct a trend offset function therefrom; the calculation formula of the trend offset function is: ; wherein, represents the function value of the suspended node in any one of the subsequent continuous periods, wherein any one of the periods refers to a target period selected from the subsequent continuous periods for calculation using the trend offset function; represents the number of related abnormal indexes of the suspended node in any one of the subsequent continuous periods, and simultaneously, any one of the periods refers to the target period as described above; refers to the same period, and the target period needs to be calculated using the trend offset function; represents the first related abnormal index in the target period; ; wherein, represents the value of the first related abnormal index in the target period, represents the value of the first related abnormal index in the previous period of the target period.

[0108] ​​If the accumulated amount of the trend offset function value in consecutive periods does not exceed the preset accumulation threshold and the change direction is reasonable, the temporary flag is eliminated and the abnormal processing action is abandoned, wherein the preset accumulation threshold refers to the upper limit value of the sum of the function values in consecutive periods set by querying existing data and expert experience; if the accumulated amount of the trend offset function value in consecutive periods does not exceed the preset accumulation threshold, and the change direction is the correct direction, which means that it presents a steady or oscillatory fluctuation, which indicates that the node cannot cause serious fluctuation out of control in this state, so the temporary flag is eliminated and the abnormal processing action is abandoned.

[0109] If the trend offset function value in consecutive periods continues to be greater than the preset function threshold and the change direction is an upward trend, the abnormal processing action is directly executed, wherein if the trend offset function value in consecutive periods continues to be greater than the preset function threshold, the accumulated amount at this time is already higher than the preset accumulation threshold, and the direction conforms to the direction of aggravation, at this time the corresponding node is considered to be in a real abnormal state, and the corresponding abnormal processing action is directly triggered for processing; the node running response strategy is constructed by integrating appropriate strategy instructions and the coded abnormal processing action execution process. The preset function threshold is set by a person skilled in the art according to actual needs.

[0110] The state fluctuation matrix is constructed in the following manner:

[0111] All related abnormal indicators belonging to the same node are normalized to obtain normalized abnormal indicators, wherein the dimensional difference of different indicators is eliminated by normalization processing, which facilitates reasonable mathematical calculation; the normalized abnormal indicators are spliced into row vectors, and the normalized abnormal indicator row vectors corresponding to all nodes are superimposed to obtain a state fluctuation matrix, wherein each column of the state fluctuation matrix represents a parameter of one dimension of the normalized abnormal indicators, and each row represents all normalized abnormal indicators corresponding to one node.

[0112] The embodiment optimizes the node synchronization, alarm mechanism, abnormal response, communication detection and strategy generation by collecting the running state data of the ARM node, and realizes an intelligent operation and maintenance monitoring method for the ARM computing power node. Compared with the existing experience, by introducing the time synchronization process of multiple nodes, the problem of clock inconsistency caused by different hardware structures of the board card is effectively solved, so that the system can adapt to the second-level response in the application scenario. By fusing the task type and resource sensitivity, the system can adjust the alarm mechanism threshold according to the actual business impact of the task, avoiding the false alarm or missed alarm situation easily produced by the traditional alarm rule. By introducing semantic content detection and skew analysis, the judgment mode of "response means communication success, otherwise communication failure" in the traditional communication detection technology is broken through, effectively alleviating the abnormal misjudgment situation and improving the accuracy of communication exception positioning. By detecting the fluctuation trend of the node index, the anti-jitter control of the computing power node running is realized, effectively avoiding the high overhead operation of the system executing abnormal response strategy due to the repeated slight fluctuation of the related index of the node, and improving the stability of the overall operation. In summary, the intelligent operation and maintenance monitoring method for the ARM computing power node significantly improves the adaptability and processing accuracy of the node abnormal information in the computing power network, and provides stronger operation adaptability and service continuity guarantee.

[0113] Embodiment 2

[0114] Please refer to Figure 2 The embodiment does not describe part in detail, see the description of embodiment 1, provide an intelligent operation and maintenance monitoring system for ARM computing power node, including:

[0115] The data acquisition module is used for collecting the running state data of the ARM computing power node and performing data cleaning to obtain node running state data.

[0116] The node synchronization module is used for performing node time synchronization processing on the node running state data to generate node synchronization running data.

[0117] The alarm optimization module is used for optimizing the alarm mechanism based on the node synchronization running data to obtain adaptive alarm rules.

[0118] The abnormality identification module is used for executing abnormal response identification on the node synchronization running data in combination with the adaptive alarm rules to output an abnormal response data queue.

[0119] The communication detection module is used for detecting the communication state of the abnormal response data queue to obtain communication state marking data.

[0120] The strategy generation module is configured to generate a node operation response strategy based on the communication state marking data and send the node operation response strategy to a preset ARM computing power node control terminal.

[0121] The above merely describes preferred embodiments of the present application and is not intended to limit the present application. Although the foregoing embodiments of the present application have been described in detail, those skilled in the art should understand that the technical solutions described in the foregoing embodiments can be modified, or some technical features can be replaced by equivalent features, without departing from the spirit and principle of the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall fall within the scope of protection of the present application.

[0122] It should be noted that, in this document, the terms "comprising", "containing", or any other variant thereof are intended to cover non-exclusive inclusion, so that a process, method, article, or apparatus that includes a list of elements not only includes those elements, but also includes other elements not explicitly listed, or inherent to such process, method, article, or apparatus. Without more limitations, the element defined by the statement "comprising a" does not exclude the presence of additional identical elements in the process, method, article, or apparatus that includes the element.

[0123] In the description of the present application, it should be understood that the terms "first", "second", etc. are only used to distinguish the description, and cannot be understood as indicating or implying relative importance.

[0124] In the description of the present application, unless otherwise specified, "a plurality of" means two or more.

[0125] In the description of the present application, "several" means one or more, and "a large number" means two or more.

[0126] In the description of the present application, the description of the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present application. In the present specification, the illustrative description of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0127] The formula in the present specification is a dimensionless value calculated from the numerical value, the formula is a formula obtained by software simulation of the most recent real situation based on a large amount of data, and the preset parameters and threshold values in the formula are set by a person skilled in the art according to the actual situation.

[0128] While embodiments of the application have been shown and described, it is to be understood that the application is not limited to the details of the embodiments described, since the scope of the application will be defined with respect to the claims and their equivalents.

Claims

1. A method for intelligent operation and maintenance monitoring of ARM computing nodes, characterized in that, include: S1. Collect the running status data of the ARM computing nodes and clean the data to obtain the node running status data; S2. Perform node time synchronization processing on node running status data, including: Obtain the board number to which the ARM computing node belongs and group the node running status data, and output the grouped node data; set the global sampling time standard for the node unit; calculate the difference between the timestamp of the grouped node data and the global sampling time standard for the node unit to obtain the time offset of the corresponding board node unit; compare the time offset with the preset time error tolerance threshold, and divide the node running status data into a subset of synchronous node data and a subset of asynchronous node data based on the comparison result; Feature extraction is performed on a subset of asynchronous node data to obtain node time features. Based on the node time features, time-series clustering is performed on the subset of asynchronous node data to generate a set of hysteresis node clusters and a set of drift trend node clusters. Historical sampling response data of different boards is obtained, and time compensation mapping is performed on the set of hysteresis node clusters based on the historical sampling response data to output a set of time-corrected node clusters. The drift period of the set of drift trend node clusters is extracted and the drift rate is calculated. Based on the drift period and drift rate, the set of drift trend node clusters is reparameterized on the time axis to obtain a set of time-reconstructed node clusters. The synchronized node data subset, the set of time-corrected node clusters, and the set of time-reconstructed node clusters are integrated to obtain node synchronization operation data. S3. Optimize the alarm mechanism based on node synchronous operation data to obtain adaptive alarm rules; S4. Combine adaptive alarm rules to perform abnormal response identification on node synchronous operation data and output an abnormal response data queue; S5. Perform communication status detection on the abnormal response data queue to obtain communication status marker data; S6. Generate a node operation response strategy based on the communication status marker data, and send the node operation response strategy to the preset ARM computing power node control terminal.

2. The intelligent operation and maintenance monitoring method for an ARM computing node according to claim 1, characterized in that, The methods for optimizing the alarm mechanism include: Obtain the running task identifier corresponding to the real-time running data of the node synchronization, and match the running task identifier with the node number to obtain the task node index set; determine the task type based on the task node index set, and construct the resource allocation index for each task type; extract the historical node data of the corresponding task type, calculate the multi-dimensional parameter difference between the resource allocation index of the historical node data and the resource allocation index of the current node, assign sensitivity weights to the resource allocation index based on the multi-dimensional parameter difference, and output the weighted resource allocation index. Interactive correlation analysis is performed on weighted resource allocation indicators to output indicator fluctuation correlation characteristics; task type anomaly judgment conditions are constructed by combining indicator fluctuation correlation characteristics and task type; a mapping relationship between task type anomaly judgment conditions and weighted resource allocation indicators is established, and anomaly correlation dataset is constructed based on the mapping relationship; anomaly subsets with overlapping tasks in the anomaly correlation dataset are identified, and alarm intensity sequences are generated based on the trigger response order of relevant indicators in the anomaly subset; adaptive alarm rules are constructed by combining relevant indicators, task type anomaly judgment conditions, and alarm intensity sequences.

3. The intelligent operation and maintenance monitoring method for an ARM computing node according to claim 2, characterized in that, The methods for constructing abnormal task type determination conditions include: Task types include cloud gaming, mobile office, and cloud phones; When the task type is cloud gaming, the abnormal judgment conditions for the task type include the short-term change in GPU utilization, the number of consecutive frame rate jitters, and the image rendering latency. When the task type is mobile office, the conditions for judging the abnormality of the task type include the rate of sudden change of CPU, memory allocation and task switching frequency. When the task type is cloud phone, the abnormal judgment conditions for the task type include system access latency, network latency jump frequency, and number of data interruptions.

4. The intelligent operation and maintenance monitoring method for an ARM computing node according to claim 3, characterized in that, The methods for identifying abnormal response execution include: Using adaptive alarm rules as the basis for judgment, identify abnormal response in node synchronous operation data and output preliminary screening of abnormal data; identify the operation behavior corresponding to the preliminary screening of abnormal data and filter the abnormal processing subset based on the operation behavior; set an abnormal danger level for each abnormal processing subset and determine the relevant abnormal indicators with an abnormal danger level higher than the preset danger level threshold as preferred abnormal indicators. Construct a dynamic sliding window to traverse all exception handling subsets, obtain all preferred exception indicators within the dynamic sliding window at the same time, and determine the preferred exception indicators as the highest processing priority; For non-optimal anomaly indicators, the distribution density and duration of the non-optimal anomaly indicators are obtained using a dynamic sliding window at the same time. The load status characteristics of the corresponding ARM computing nodes are collected in real time. The suboptimal anomaly score is calculated by combining the values ​​of distribution density, duration and load status characteristics. The non-optimal anomaly indicators are sorted based on the suboptimal anomaly score to obtain the suboptimal anomaly queue. The optimal anomaly indicators and the suboptimal anomaly queue are integrated to obtain the anomaly response data queue.

5. The intelligent operation and maintenance monitoring method for an ARM computing node according to claim 4, characterized in that, The methods for performing communication status detection include: Obtain the communication path set of the ARM computing node corresponding to the abnormal response data queue; extract the node response data of each communication path in the communication path set, identify the semantic content of the node response data, and construct a node semantic vector matrix based on the semantic content; perform structural integrity analysis on the node semantic vector matrix, and filter the initial abnormal response data packets based on the structural integrity analysis results; Identify the request instructions corresponding to the response data of other nodes, and query the original included fields based on the request instructions; calculate the mismatch rate by comparing the number of existing fields with the original included fields, and perform two-category filtering on the response data of other nodes based on the mismatch rate and a preset threshold to obtain two types of abnormal response data packets; Extract the resource occupancy rate of the response data of the remaining nodes and construct an abnormal field occupancy vector; calculate the standard deviation of all elements in the abnormal field occupancy vector, perform skewness analysis on the abnormal field occupancy vector based on the standard deviation, and output the skewness coefficient; if the skewness coefficient does not meet the preset standard coefficient range, perform three-class filtering on the response data of the remaining nodes to obtain three types of abnormal response data packets; mark the communication abnormal status of each filtered abnormal response data packet to obtain communication status marking data.

6. The intelligent operation and maintenance monitoring method for an ARM computing node according to claim 5, characterized in that, The methods for marking communication anomalies include: If the structural integrity analysis result indicates that the structure is incomplete, then the initial abnormal response data packet at this time will be judged as a structural damage abnormal state; if the mismatch rate is higher than the preset threshold, then the second type of abnormal response data packet at this time will be judged as a semantic content abnormal state; if the skewness coefficient does not conform to the preset standard coefficient range, then the third type of abnormal response data packet at this time will be judged as a resource writing abnormal state.

7. The intelligent operation and maintenance monitoring method for an ARM computing node according to claim 6, characterized in that, The methods by which the generating node executes the response strategy include: Extract communication status tags from communication status marker data, match the communication status tags and related abnormal indicators with a pre-set abnormality resolution strategy database, and generate appropriate strategy instructions. Extract the relevant abnormal indicator fluctuation value of any node within a continuous preset monitoring period, and construct a state fluctuation matrix based on the relevant abnormal indicator fluctuation value; set a reasonable fluctuation range, and count the number of times the fluctuation value of all relevant abnormal indicators of each node in the state fluctuation matrix exceeds the reasonable fluctuation range; if the number is greater than a preset frequency threshold, execute the abnormal handling action directly; otherwise, mark the corresponding node as temporarily suspended, and record the trend change of the relevant abnormal indicators of the temporarily marked node in subsequent continuous periods, and construct a trend offset function; if the accumulated amount of the trend offset function value in the continuous period does not exceed the preset accumulated amount threshold and the change direction is correct, remove the temporary suspension mark and abandon the execution of the abnormal handling action; if the trend offset function value is continuously greater than the preset function threshold in the continuous period and the change direction is upward, execute the abnormal handling action directly; integrate appropriate strategy instructions and the coded abnormal handling action execution process to construct a node operation response strategy.

8. The intelligent operation and maintenance monitoring method for an ARM computing node according to claim 7, characterized in that, The methods for constructing the state fluctuation matrix include: Normalize all relevant abnormal indicators belonging to the same node to obtain normalized abnormal indicators; concatenate the normalized abnormal indicators into row vectors, and superimpose the row vectors of normalized abnormal indicators corresponding to all nodes to obtain the state fluctuation matrix.

9. An intelligent operation and maintenance monitoring system for ARM computing nodes, used to implement the intelligent operation and maintenance monitoring method for ARM computing nodes as described in any one of claims 1-8, characterized in that, include: The data acquisition module is used to collect the operating status data of the ARM computing nodes and perform data cleaning to obtain the node operating status data. The node synchronization module is used to perform node time synchronization processing on node running status data and generate node synchronization running data. The alarm optimization module is used to optimize the alarm mechanism based on node synchronous running data to obtain adaptive alarm rules; The anomaly detection module is used to identify anomalies in node synchronization data by combining adaptive alarm rules and output an anomaly response data queue. The communication detection module is used to detect the communication status of the abnormal response data queue and obtain communication status marker data. The strategy generation module is used to generate node operation response strategies based on communication status marker data and send the node operation response strategies to the preset ARM computing power node control terminal; the modules are connected to each other via wired and / or wireless means.

Images